Comparative evaluation of thyroid nodule ultrasound segmentation models

Objectives To compare the performance of existing thyroid nodule ultrasound segmentation models and provide evidence for their clinical application. Methods The models were trained and evaluated under standardized conditions on two datasets: the TN5000 public dataset (5,000 images), which was split into training and internal validation sets, and an institutional dataset (1,452 images) used as an independent external test set. Results TNet achieved the best overall performance (external Dice = 0.812), and LGFFM showed higher recall. All models exhibited a consistent and substantial decline in performance over cross-center validation, with non-thyroid-specific models performing worst. Conclusions Existing models have limited cross-center generalizability, and TNet performed best among the evaluated models in this experimental setting, while broader clinical adoption would require prospective multicenter validation. Clinical trial number Not applicable.